Histologic Image-Based Classifier for Predicting Outcome of ER+ Breast Cancers.
Ajay N Basavanhally, Shridar Ganesan, Michael Feldman, John Tomaszewski, Anant Madabhushi. Rutgers, The State University of New Jersey, Piscataway; The Cancer Institute of New Jersey, New Brunswick; Hospital of the University of Pennsylvania, Philadelphia
Background: The Oncotype DX molecular assay uses expression levels of 21 genes to produce a Recurrence Score (RS) and help determine prognosis for ER+ breast cancer (BCa) patients. However, it suffers from translational limitations (e.g. time and cost per test). It has been shown that prognostic information in ER+ BCa is reflected in histopathology, but manual analysis suffers from inter-pathologist variability. We address these shortcomings via the Image-based Risk Score (IbRiS), a computerized decision support system that predicts disease outcome using only digitized images of H & E stained biopsy samples.
Design: IbRiS is based on the idea that differences in outcome are reflected by variations in tissue architecture. Hence we extract quantitative image features relating the spatial arrangement of BCa nuclei. Nuclei are first automatically detected using a color deconvolution scheme to isolate hematoxylin stain. Nuclear centroids are treated as vertices for the construction of 3 graphs, from which 50 features are extracted for each image. The data is subsequently projected down into a reduced space via a machine learning scheme called Graph Embedding (Fig. 1).
Results: A total of 300 histopathology images were obtained from 50 patients with corresponding B-R grades and RS values. In Fig. 1, each data point is an image in the reduced space and B-R grade labels represent disease outcomes. The labels suggest that the data manifold models risk of disease progression on a smooth continuum. By “unwrapping” the manifold, prognostic thresholds could be constructed to guide therapy similar to RS. In addition, cross-validation shows that the image-based features correlate B-R grade and RS values (low vs. high) at 0.90 ± 0.02 accuracy.
Conclusions: IbRiS is a novel companion prognostic tool that automatically detects BCa nuclei, quantifies the spatial arrangement of those nuclei, and reveals the underlying manifold that can stratifies outcome. By using only digitized histopathology, IbRiS offers a fast, inexpensive, and highly accessible decision support system for ER+ BCa prognosis.
Wednesday, March 2, 2011 9:30 AM
Poster Session V # 15, Wednesday Morning